
==== Front
Proc Natl Acad Sci U S A
Proc Natl Acad Sci U S A
PNAS
Proceedings of the National Academy of Sciences of the United States of America
0027-8424
1091-6490
National Academy of Sciences

38527192
202314918
10.1073/pnas.2314918121
research-articleResearch ArticleneuroNeuroscience424
Biological Sciences
Neuroscience
Subcallosal cingulate deep brain stimulation evokes two distinct cortical responses via differential white matter activation
Seas Andreas a b 1 https://orcid.org/0000-0003-0624-1254

Noor M. Sohail a c 1 https://orcid.org/0000-0003-4139-888X

Choi Ki Sueng d e https://orcid.org/0000-0002-6446-0677

Veerakumar Ashan e
Obatusin Mosadoluwa d e
Dahill-Fuchel Jacob d
Tiruvadi Vineet e
Xu Elisa d
Riva-Posse Patricio e
Rozell Christopher J. f
Mayberg Helen S. d e https://orcid.org/0000-0002-1672-2716

McIntyre Cameron C. a b c
Waters Allison C. d e 2 https://orcid.org/0000-0002-6374-3712

Howell Bryan bryan.howell@duke.edu
a c 2 3 https://orcid.org/0000-0002-3329-8478

aDepartment of Biomedical Engineering, Duke University, Durham, NC 27708
bDepartment of Neurosurgery, Duke University, Durham, NC 27708
cDepartment of Biomedical Engineering, Case Western Reserve University, Cleveland, OH 10900
dNash Family Center for Advanced Circuit Therapeutics, Icahn School of Medicine at Mount Sinai, New York, NY 10029
eDepartment of Psychiatry and Behavioral Sciences, Emory University School of Medicine, Atlanta, GA 30329
fSchool of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA 30332
3To whom correspondence may be addressed. Email: bryan.howell@duke.edu.
Edited by Donald Pfaff, Rockefeller University, New York, NY; received September 1, 2023; accepted February 20, 2024

1A.S. and M.S.N. contributed equally to this work.

2A.C.W. and B.H. contributed equally to this work.

25 3 2024
2 4 2024
25 9 2024
121 14 e231491812101 9 2023
20 2 2024
Copyright © 2024 the Author(s). Published by PNAS.
2024
https://creativecommons.org/licenses/by-nc-nd/4.0/ This article is distributed under Creative Commons Attribution-NonCommercial-NoDerivatives License 4.0 (CC BY-NC-ND).

Significance

This study characterizes the cortical evoked response to selective neural pathway perturbations in the subcallosal cingulate of patients with depression undergoing deep brain stimulation. Using structural imaging, diffusion tensor tractography, and field-cable modeling, we discern which white matter tracts are stimulated by specific stimulation settings and correlate this activation with associated scalp electroencephalography. We demonstrate unique circuit responses corresponding to selective perturbation of white matter tracts within the therapeutic target region, providing validation for using pathway activation modeling to predict neural responses to brain stimulation for treating depression. We demonstrate how evoked responses may serve as fiber-specific “fingerprints” and provide direct readouts from the brain to improve the programming of the implanted device.

Subcallosal cingulate (SCC) deep brain stimulation (DBS) is an emerging therapy for refractory depression. Good clinical outcomes are associated with the activation of white matter adjacent to the SCC. This activation produces a signature cortical evoked potential (EP), but it is unclear which of the many pathways in the vicinity of SCC is responsible for driving this response. Individualized biophysical models were built to achieve selective engagement of two target bundles: either the forceps minor (FM) or cingulum bundle (CB). Unilateral 2 Hz stimulation was performed in seven patients with treatment-resistant depression who responded to SCC DBS, and EPs were recorded using 256-sensor scalp electroencephalography. Two distinct EPs were observed: a 120 ms symmetric response spanning both hemispheres and a 60 ms asymmetrical EP. Activation of FM correlated with the symmetrical EPs, while activation of CB was correlated with the asymmetrical EPs. These results support prior model predictions that these two pathways are predominantly activated by clinical SCC DBS and provide first evidence of a link between cortical EPs and selective fiber bundle activation.

deep brain stimulation
depression
evoked potentials
biophysical modeling
connectomics
HHS | NIH | National Institute of Mental Health (NIMH) 100000025 R01MH102238 Cameron C McIntyre HHS | NIH | National Institute of Neurological Disorders and Stroke (NINDS) 100000065 UH3NS103550 Patricio Riva PosseChristopher J RozellHelen S. Mayberg
==== Body
pmcSubcallosal cingulate (SCC) deep brain stimulation (DBS) is an evolving surgical treatment option for people with severe, unremitting depression with response rates of 23 to 92% across 14 clinical studies to date (1–3). The SCC is metabolically overactive in people with depression, and high-frequency (130 Hz) DBS was initially hypothesized to normalize this overactivity through a reversible lesioning effect (4). The target for SCC DBS has shifted over time from SCC white matter to the crossroads of four specific bundles, including the cingulum bundle (CB), forceps minor (FM), uncincate fasciculus (UF), and connections from the frontal pole (FP) to subcortical regions (1, 5–7). Implanting the lead more precisely at the confluence of these four pathways with tractographic guidance has consistently improved responses by over 70% in recent cohorts, but it is still unclear which of these target bundles need to be activated and in what proportions to achieve an optimal outcome (8–10). The present study addresses a methodological obstacle to conducting this critical inquiry.

Electrophysiological readouts from the brain provide insights into network engagement resulting from white matter activation (11–17). Intracranial electroencephalography (iEEG) is one approach to map engagement of gray matter with millimeter spatial resolution (18). iEEG was used by Adkinson et al. (19) to map the cortical networks engaged by various white matter pathways, suggesting the involvement of striatal, amygdala, and prefrontal cortical nodes depending on pathway activation. However, iEEG is not a scalable strategy in large patient populations due to its cost and surgical risks.

Scalp EEG is an alternative for mapping network-wide brain engagement. Scalp EEG has low spatial resolution (5 to 9 cm) compared to iEEG, but its millisecond temporal precision and noninvasive application make it a scalable solution (20). Waters et al examined the reliability of cortical responses to SCC DBS (21). Unilateral activation of white matter by SCC DBS at the therapeutic contact produced a symmetrical EP spanning frontal and temporal cortical sensors. This stimulation-induced EP was similar across patients and time, persisting for up to 14 mo after DBS electrode implantation, and the threshold amplitude for evoking this EP differed across contacts. The same EP has also been observed independently in another SCC DBS cohort (22), but it is still unclear which specific fiber bundles drive this response and if there are differentiable cortical evoked responses for the activation of different fiber tracts.

The goal of this study was to characterize the EPs of specific white matter tracts targeted in SCC DBS. Our latest biophysical modeling of SCC DBS predicts that FM and CB are activated in the greatest proportions in responders, so this study focuses on these pathways (9). We used detailed connectomic modeling of SCC DBS (23, 24) to identify and optimize selective settings for activating either CB or FM in isolation in seven depression patients who were responders. We recorded EEG responses to both selective stimulation settings and the patient’s therapeutic stimulation setting but at 2 Hz. The study aimed to 1) compare the EPs elicited by selective activation of one pathway to nonselective coactivation of both pathways, 2) identify which features of the EPs were unique to each pathway, and 3) determine whether the EP features scaled with the percentage of fiber recruitment. Preliminary results of this study were presented separately in conference proceedings (25).

Results

MRI-based head models of SCC DBS were constructed for seven SCC DBS patients using the latest advances in patient-specific biophysical modeling of DBS (Fig. 1) (9, 23). These connectomic DBS models identified optimal selective stimulation settings for each patient, which were applied to elicit cortical responses, and the resulting EPs were recorded with high-density EEG. All stimulation settings tested were unilateral (Fig. 2D). Selective settings included model-optimized settings and clinical settings that activated only one pathway with one lead. Nonselective settings included clinical settings in patients where both pathways were coactivated and model-optimized settings that were almost selective but still had some coactivation. A total of 118 different EPs were recorded across all subjects. Of these, 83 had perceivable signals above the noise floor, whereas the other 35 were classified as having only noise. Among the 83 EPs with signals, 40 were for selective stimulation of CB, 3 for selective stimulation of FM, and 40 for nonselective activation of both pathways.

Fig. 1. Optimizing selective activation of white matter targets using connectomic modeling. Pathways were reconstructed using tractography for (A1) forceps minor (FM) and (A2) each cingulum bundle (CB). The electric field generated by SCC DBS was simulated for each patient and applied to (B1) FM and (B2) CB. Three types of settings were tested in each patient. (C1) One setting used the clinical settings but at 2Hz. (C2) The percentage of pathways activated by the clinical settings for all programmable voltages. (D1) One setting was optimized by the field-cable model to selectively activate the left or right CB. (D2) The pathways activated by the selective CB setting. (E1) One setting was optimized by the model to activate FM. (E2) The pathways activated by the selective FM setting. The left CB is shown as a representative example for the CB pair. Top-right insets in B1–E1 and B2 show the virtual DBS lead with its respective programmed settings. All stimulation was applied at 2Hz. The tracts shown in C1–E1 are activated at the respective applied voltage shown in C2–E2 (dashed line).

Fig. 2. Simultaneous recordings of evoked potentials (EPs) with 256-sensor EEG. (A) Two selective model-based settings and one nonselective setting were administered sequentially while the clinical settings were temporarily halted. EEG was recorded during each stimulation setting at 2 Hz. (B) Stimulation was applied over four epochs with increasing stimulation voltages from V1 to V4. The voltage at one EEG sensor is shown for all four epochs. The Inset shows a zoomed view of the voltage–time series with four stimulation artifacts (pink rectangles). (C) Average EP in the fourth epoch at five sensors. Time series were segmented and averaged with respect to the onset of the stimulus artifact. Top–down view of all 256 sensors (black dots) projected onto a circle representing the head surface with the triangle pointing anterior. Distributions of the interpolated voltages across the head surface are shown at four time points. Positive voltage = red, negative voltage = blue. (D) Theoretical pathway activations for all the unilateral settings tested across all patients. Selective settings included model-optimized settings that activated only CB (filled yellow circles) or FM (filled red circles) and clinical settings that activated only CB (unfilled yellow circles). Nonselective settings included model-optimized settings that were almost selective but with some coactivation (filled blue circles) and clinical settings that coactivated both pathways (unfilled blue circles). No clinical settings activated only FM.

Selective SCC DBS Generated Two Distinct Cortical Responses.

Selective activation of each target pathway generated distinct EPs. CB-selective stimulation produced an asymmetrical EP (Fig. 3A) concentrated in the hemisphere of stimulation in 27 of the 40 CB-selective EPs. This asymmetrical EP had a voltage maximum at 30 ms with a left-to-right dipole inversion from 60 ms onward. Eleven of the 40 CB-selective EPs were symmetrical, whereas the remaining 2 EPs could not be classified as symmetrical or asymmetrical between three separate raters. In contrast, nonselective stimulation produced a ~120 ms symmetrical EP in 32 out of 40 nonselective stimulation cases. This symmetrical EP was characterized by a positive voltage maximum that started at the bilateral frontal sensors at ~30 ms, moved to the central midline sensors by ~60 ms, and then moved to the posterior sensors by ~90 ms. The symmetrical positive wave was followed by a similar but shorter negative waveform from 90 to 120 ms, which was characteristic of a dipole inversion. Only 7 of the 40 nonselective EPs were asymmetrical, and 1 EP could not be classified. FM activation produced an EP similar to that of nonselective stimulation in all 3 FM-selective cases (Fig. 3B). All subjects had a combination of asymmetrical and symmetrical EPs except for subject 4, who had only symmetrical EPs. The asymmetrical and symmetrical EPs were statistically separable based on their left–right symmetry (Fig. 3C), but subject as a covariate did not predict the symmetry of the recordings. Therefore, we posit that FM activation is the primary driver of the symmetrical EP.

Fig. 3. Selective activation of each pathway by SCC DBS generates distinct EPs. EPs for (A1) nonselective stimulation and (A2) selective CB stimulation in a representative subject. (A3) An example of pairwise symmetry between left and right sensors for a CB-selective EP. Programmed settings are shown to the right of each respective modeled pathway activation. EPs for the (B1) nonselective stimulation and (B2) CB-selective stimulation in another subject. (B3) An example of pairwise symmetry for an FM-selective EP. The lightning bolt marks the side of stimulation. (C) Distribution of symmetry scores in 30 ms time bins across all unilateral settings tested for all patients. Selective settings included model-optimized settings that activated only CB (filled yellow circles) or FM (filled red circles), and clinical settings that activated only CB (unfilled yellow circles). Nonselective settings included model-optimized settings that were almost selective but with some coactivation (filled blue circles), and clinical settings that coactivated both CB and FM (unfilled blue circles). No clinical settings activated only FM. The quartiles of the distributions are summarized by the I-bars. Kruskal–Wallis tests were conducted to test for equality between the different distributions (α = 0.05). The single and triple asterisks denote a P-value <0.05 and <0.001, respectively.

The Symmetrical EP May Overshadow the Asymmetrical EP.

Symmetrical EPs were larger and more spatially focal than asymmetrical EPs (Fig. 4). Selective CB stimulation produced the smallest EP, with a median magnitude of 3.32 µV, and the sensor where the voltage peaked varied across patients (Fig. 4 A or B). Voltage extrema of these CB asymmetrical EPs spanned frontal, temporal, and central sensors. By comparison, selective FM activation produced a much larger EP with a median magnitude of 12.5 µV (Fig. 4C), while coactivation of FM and CB produced an EP with a median magnitude of 7.88 µV (Fig. 4 D1), suggesting destructive interference between the CB selective EP and FM selective EP. The voltage extrema of the symmetrical EPs were concentrated in the frontopolar sensors above the orbits of the eyes. The symmetrical EPs generated from CB activation had characteristics of both CB and FM activation. The symmetrical CB EPs were small like the asymmetrical EPs (Fig. 4 A or B), and their voltage extrema were concentrated in frontopolar sensors like FM EPs (Fig. 4 B or D1). The magnitude of these less common symmetrical CB EPs correlated strongly with CB activation (Fig. 4 B, Right panel); but overall, EP magnitudes were poorly predicted by the percentage of pathway activations.

Fig. 4. Asymmetrical EPs are smaller and more diffuse than symmetrical EPs. (A) Selective CB activation produced (27 of 40) asymmetrical EPs with voltage maxima (Bottom panels) at 30 ms distributed across the scalp for different hemispheres and patients. There was no correlation between CB activation and voltage maxima. (B) Selective CB activation also produced a minority (11 of 40) of symmetrical EPs that correlated with the voltage maxima. (C) Selective FM activation produced large symmetrical EPs in all cases (3 of 3) that correlated with the respective voltage maxima. (D1) Coactivation of CB and FM produced (32 of 40) symmetrical EPs similar to those of selective FM activation. Percent activation of each pathway poorly predicted the voltage maxima for (D2) nonselective symmetrical EPs and (D3) nonselective asymmetrical EPs. The voltage magnitude maximum was calculated across all sensors at time = 30 ms (p30). The lightning bolt marks the side of stimulation. Linear regressions were fit using the least squares method with P < 0.05 being significant.

We posited that the smaller CB EP may be concealed by the larger FM EP. To explore this concept, we calculated the putative sources of nonselective EPs (Fig. 5). If the CB EP was concealed, then the signal would be largely symmetric from FM activation but also contain secondary sources that were more lateralized. On average, the symmetrical non-selective EPs stemmed from sources within the bilateral frontal polar cortex. However, there were also additional sources that were predominately on the side of stimulation (SI Appendix, Fig. S1), and the most common asymmetrical source was identified in the ventromedial prefrontal cortex (vmPFC) (Fig. 5A2 and D). This was consistent with the CB EP amplitudes staying elevated from 30 to 60 ms at frontopolar sensors (Figs. 3A and 6A), while voltages at these same sensors for nonselective EPs were near zero as the dipole inverted (Figs. 3B and 6D). Therefore, there was a time window where the evoked activity due to CB activation was more apparent and not concealed. The degree of this revealed asymmetry in the symmetrical nonselective EPs at 60 ms (Fig. 5A3) was larger when the percentage of CB activation was 2 to 3 times greater than that of FM activation (i.e., the ratio of %FM to CB% was <1) (Fig. 5C).

Fig. 5. Nonselective symmetrical EPs conceal unilateral source activity. (A1) Representative example of a symmetrical nonselective EP generated from coactivation of the left CB and FM. The lightening bold shows the side of stimulation. (A2) Dipolar sources localized from the EP in A1. activity in the FP, ventrolateral prefrontal cortex (vlPFC), and vmPFC are highlighted. Positive (red) and negative (blue) current source densities reflect current sources and sinks, respectively. (A3) Source activity in the vmPFC ipsilateral or contralateral to the side of stimulation. The gray line highlights the asymmetry of amplitudes at 60 ms. L = left, R = right. (B) Pathway percentages activated for all symmetrical nonselective EPs at their maximum stimulation amplitude. (C) Ratio of ipsilateral to contralateral source activity in the vmPFC versus the ratio of FM to CB activation. (D) Average normalized source activity across all the nonselective EPs at their maximum amplitude.

Fig. 6. Structural connectivity underlying selective and nonselective EPs. (A) Direct selective activation of the CB generates both orthodromic and antidromic action potentials that invade the ipsilateral medial frontal cortex (mFC), the FP, or the vmPFC. The SCC is shown in blue. (B) Selective CB activation may also activate callosal-striatal collaterals generating APs that symmetrically invade both frontal cortices or the posterior part of FM connecting the two SCC. The striatum (Str) is shown in green. (C) Selective activation of FM directly activates the body of the callosal axons and generates antidromic and orthodromic APs that invade FP or vmPFC. (D) Coactivation of CB and FM. The gray curves depict interhemispheric connections. Insets in each panel show the time series at the left frontopolar sensor (Fp1, purple) and right frontopolar sensor (Fp2, green). All EPs are shown at 30 ms (vertical dashed line). The lightning bolt marks the side of stimulation.

Discussion

Electrophysiological readouts from the brain are helping to standardize the clinical administration of SCC DBS. Intracranial recordings from the implanted lead show promise for tracking a patient’s therapeutic recovery (10), and whole-brain EPs may provide a reliable assay for confirming white matter and network engagement. We show that the cortical EP evoked by SCC white matter activation is not an unimodal signal, but rather a mixed signal dominated by FM recruitment. We posit that FM activation generates a relatively large signal from synchronous neural activity, overshadowing smaller asynchronous neural activity stemming from CB activation. Each pathway’s respective EPs are consistent with the pathway’s underlying connectivity (26).

The symmetrical EP had features consistent with FM’s structural connectivity. FM makes up the frontal part of the corpus callosum with its ventral portion projecting to the frontal polar and ventral frontal cortices (27–30). The amplitude of the primary positive peak in the symmetrical EP peaked around 30 ms (Fig. 6 B–D) with putative neural sources in the vmPFC (Fig. 5 A2 and D), implicating modulation of the default mode network, consistent with recent PET findings (31). The first primary peak of EPs is believed to arise from direct activation of postsynaptic neurons (16). Therefore, the primary positive peak of the symmetrical EP may arise from bilateral activation between the SCC and vmPFC (Fig. 6 C and D) (21, 22). Cellular recordings reveal that relatively large positive surface potentials can arise from synchronous depolarization of superficial (layer 2/3) pyramidal neurons, generating a current sink (negative potentials) close to the soma and an opposite source (positive potentials) closer to the cortical surface. Additionally, positive surface potentials can arise from simultaneous depolarization of the soma and dendrites of deeper (layer 5) pyramidal neurons, producing a calcium wave in the dendrites and a dipole pointing toward the cortical surface (32, 33). Later secondary peaks of the EP likely arise from multisynaptic connections between remote cortical and subcortical regions, which facilitate slower waves mediated by GABA neurotransmitters (16, 34).

The features of the asymmetrical EP were more consistent with CB’s connectivity. Whereas FM is composed of homogenous fibers reciprocally connecting large cortical areas, CB is a diverse collection of both short- and long-range projections to cortical and subcortical regions, including connections between the orbitofrontal cortex, frontopolar cortex, anterior cingulate cortex (ACC), amygdala, and thalamus (35–37). Therefore, CB activation is expected to generate more latency and dipole scatter compared to FM, leading to a more spatially distributed EP (Fig. 4A). In contrast, FM activation likely causes simultaneous depolarization of the soma and dendrites of deep pyramidal neurons, triggering calcium waves and resulting in a large EP (33). This is consistent with the FM selective EP being 2 to 3× larger than the CB selective EP (Fig. 4). This could also explain why the nonselective EP resembled the FM selective EP because the FM sources dominate the signal. However, at certain time points in the nonselective EP, CB-evoked activity is elevated with respect to FM-evoked activity, revealing underlying CB sources (Fig. 5).

Selective CB activation also produced symmetrical EPs in 5 out of 7 subjects in a minority (11/46) of the cases. Symmetrical CB EPs may originate from the activation of FM striatal collaterals (Fig. 6B) (27). A minority of layer 5 callosal fibers send collaterals to the contralateral striatum in humans, and these callosal cortico-striatal collaterals could course through the rostral–caudal CB from the frontopolar regions to the striatum (37, 38). This means activating the lateral aspects of the CB could simultaneously activate these smaller callosal collaterals. Another possibility is that our model lacked tracts in the FM that connect the bilateral SCC (Fig. 6B). These FM fibers near the rostrum abut and potentially overlap with CB tracts, so the posterior aspects of FM may have been missed in the tractographic reconstructions.

Direct brain readouts are needed in the clinic to confirm target engagement. Our results show that tract activations could be differentiated based on unique features of their respective EP, and with an understanding of a pathway’s “fingerprint,” it is possible to classify pathway perturbations using machine learning. Model guidance would steer the generation of training examples of selective and nonselective EPs, and a time series neural net, for example, could be used to classify which tracts are activated based on test EP readouts. Additionally, EEG pattern recognition could be used to test and confirm energy-efficient stimulation settings. The therapeutic contact could be used to define a reference EP, and more energy-efficient settings could be tested against this reference with the goal of matching an EP blueprint (25).

While this study focused on CB and FM, there is still no consensus among modeling studies on which tracts need to be activated to achieve an optimal response to SCC DBS. CB and FM are the primary sources of variation in the direct neuronal response to SCC DBS, but UF and frontopolar (FP) connections are also recruited in some responders (7, 9). CB activation alone may be sufficient to promote a rapid antidepressant response to stimulation (39, 40). Coactivation of FM and UF with CB may further augment positive effects (41, 42). However, excess stimulation of FM beyond a certain critical percentage may counteract the positive effects (39). A critical methodological barrier to testing these model-based finding has been a lack of an objective readout to confirm each tract’s engagement. This study demonstrates the predictive power of using pathway activation modeling to steer selective pathway recruitment, and pathways-specific EPs provide a new direction for testing new connectomic-based hypotheses.

This work excluded UF and FP connections from our analysis. UF and FP were omitted because they were only recruited in small proportions (<5%) in this cohort (9). Additionally, UF and FP could not be activated without first activating large proportions of CB or FM given their close proximity to the lead. Moreover, a significant proportion (>5%) of CB or FM activation was generally required to generate an EP above the noise floor, so we assumed that sparse recruitment (<5%) of UF or FP would be concealed within the noise floor. Finally, EEG experiments were conducted using a stimulus frequency of 2 Hz, whereas clinical therapeutic stimulation is delivered at 130 Hz. At stimulus frequencies greater than 10 Hz, EPs are obscured by stimulation artifacts, making it unclear how the EP evolves over time at higher stimulation frequencies. That said, biophysics dictates that the same white matter profile is directly activated at both 2 and 130 Hz, implying that low-frequency EPs still provide valuable information about the engaged white matter and the propagation of signals within the broader depression network.

Materials and Methods

This study evaluated patient-specific pathway-selective connectomic DBS settings in seven patients with severe, chronic depression. All patients provided written informed consent to participate in the study. The protocol was in accordance with the Declaration of Helsinki. The protocol was approved by the Institutional Review Boards at Emory University, Georgia Institute of Technology and the Icahn School of Medicine. Additionally, it received approval from the US Food and Drug Administration under a physician-sponsored Investigational Device Exemption (IDE G130107) and was monitored by the Emory University Department of Psychiatry and Behavioral Sciences Data and Safety Monitoring Board. Patients had bilateral DBS leads (Medtronic Model 3387) with the Activa PC+S DBS system (Medtronic PLC, Minneapolis, MN). The inclusion and exclusion criteria for this patient selection can be found in ref. 10.

Patient Imaging.

Medical images were collected for surgical planning and modeling. Preoperative structural MRI was collected using a Siemens 3 T Trio Tim scanner (Siemens Medical Solutions, Malvern, PA): maximum gradient = 40 mT/m, 176 slices, 1 × 1 × 1 mm3, TR/TE = 2,600/3.02 ms, flip angle = 8°). Diffusion-weighted MRI was collected using high-angular resolution imaging with a 32-channel coil, 11 nondiffusion weighted images (b = 0), and 128 noncollinear directions: b value = 1,000 s/mm2, 2 × 2 × 2 mm3, 64 slices, TR/TE = 3,292/96 ms, multi-band acceleration factor = 3 (43). Postoperative computed tomography (CT) was collected 3 wk after surgery using a LightSpeed 16 scanner (GE Medical System, 0.46 × 0.46 × 0.65 mm3 voxel size). MR and CT images were processed using the FMRIB Software Library (FSL, https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/) and the Analysis of Functional Neuroimages toolbox (AFNI, https://afni.nimh.nih.gov/) (44, 45). Details on the image processing methods are described previously by Howell et al. (9, 44). White artifacts of the DBS leads were localized in the postoperative CT using the PaCER algorithm (46). All images were registered to the patient’s preoperative T1 MRI anatomical space for model evaluation.

Tractographic Pathway Reconstructions.

To capture intersubject differences in connectivity, FM and CBs were reconstructed (Fig. 1A) from each patient’s diffusion-weighted MRI using probabilistic tractography (47). Diffusion parameters were calculated using FSL’s bedpost, and streamlines were generated in the diffusion MRI toolkit, Camino, using the FACT “ball-and-stick” tracking algorithm (48, 49). One hundred streamlines were seeded per voxel within a region of interest (~30 K voxels) defined as the union of eight 30 mm spheres surrounding all eight electrodes (four electrodes/lead). Streamlines with a curvature >60° over 5 mm arc or total arc length <40 mm were rejected. For this study, we focused only on the ventral portion of FM below the anterior-most point (i.e., the frontal tip) of the genu. Details on the tractography are summarized by Howell et al. (9). Streamlines were all modeled using a fiber diameter (axon diameter + myelin thickness) of 2 µm. While 70% of frontal white matter is myelinated with fiber diameters <0.4 µm, axons this small cannot be directly activated even at the highest available clinical DBS stimulation intensities (10 mA) (50–52). Larger axons with a diameter >2 µm comprise <2% of frontal tracts, but these axons are excitable at clinical stimulation settings and thereby likely to constitute the majority of axonal activations by SCC DBS (9, 51).

Field-Cable Modeling.

Patient-specific connectomic models are the combination of a field model of electric potentials generated by the DBS lead and cable-based model estimates of axonal activation. Previous studies have described the construction of these models in detail, and the process is summarized in Fig. 1 (9, 23, 24, 53).

Volume conductor models of the DBS electric field were built using the finite element method (FEM) analysis software, COMSOL (v5.1). The head volume included two DBS leads (Medtronic Model 3387), encapsulation tissue (0.5 mm thickness, σ = 0.1 S/m) representing a glial scar and bulk tissue (12 cm cube, σ = 0.2 S/m) representing the head. Tissue volumes were purely resistive. Locations of the DBS leads were calculated from the postoperative CT. Then, the electric potential generated by DBS was calculated by solving Laplace’s equation,[1] ∇·σ·∇Φ,

where σ is the tissue conductivity and Φ is the scalar fields of electric potentials. Φ was applied to each streamline (Fig. 1B) modeled as an axon with a fiber diameter of 2 µm, and activation thresholds for all axons were calculated using a driving-force predictor (24, 54). The driving force algorithm reduces solve times for cable-based models drastically from hours to milliseconds. This enables the evaluation of thousands of model solutions to identify the globally optimal selective stimulation parameters for FM and CB (Fig. 1).

Model-Based Optimization of Selective Pathway Activation.

Connectomic models were used to optimize selective activation of each pathway (Fig. 1). Selective stimulation settings activated one target pathway without coactivation of the other pathway. These selective settings were chosen among 774 options programmable with the implanted Medtronic Model 3387 lead. The best selective setting for each target pathway was the one that maximized the normalized area under the respective percent target versus percent nontarget curve. A normalized area of 1 indicates 100% activation of the target pathway with no activation of other pathways, while 0 means no activation of the target pathway (55). For selective CB stimulation (Fig. 1D), CB was the target and FM was the nontarget. Similarly, for FM selective stimulation (Fig. 1E), FM was the target and CB was the nontarget. We also tested unilateral clinical settings but at 2 Hz (instead of the clinically used 130 Hz) as a standard for comparison. Some clinical settings activated both CB and FM and were thereby classified as nonselective (Fig. 1C). Other clinical settings activated only CB and were thereby classified as selective. No bilateral settings, either clinical or model-based, were tested in this study. It should be noted that despite model optimizations, perfect selective activation of CB or FM was not possible in all patients. However, these nonselective (almost selective) settings were still tested because they covered cases where CB activation was much greater than FM and vice versa.

Electroencephalography (EEG) Recording.

High-density EEG was performed during clinical visits—anywhere from 16 wk to 1 y after the chronic stimulation settings were initiated. For EEG recordings, therapeutic stimulation was discontinued an hour before the experiment. Patients were fitted with a 256-channel Magstim-EGI Sensor Net (Whitland, United Kingdom) and seated with an adjustable chinrest in a climate-controlled environment. EEG was acquired at a sampling rate of 1 kHz. Electrodes were checked after each recording session, and their impedances were maintained below 50 kΩ by applying saline solution. More details on the EEG protocol can be found in Waters et al. (21). EEG recordings were performed during unilateral 2 Hz stimulation at the therapeutic setting (contact and pulse width) and FM-selective and CB-selective settings. The therapeutic settings were applied at 2, 3, 4, and 5 V. Selective settings were applied at 2, 4, 6, and 8 V. These voltage ranges were broad enough to evoke an EP by the maximum stimulation amplitude. Each stimulation block lasted approximately 3 min.

EEG Preprocessing.

EEG was high-pass filtered with a cutoff at 0.4 Hz and notch filtered at 60 Hz. The SD of all channels within an experiment was recorded, and channels with values exceeding the 95% bounds of a Gaussian describing these data were removed. Each sensor was rereferenced to a virtual reference channel, defined as the instantaneous mean voltage across all good recording channels. Timestamps of stimulation artifacts were identified by first computing a moving variance of the FCz channel, as well as high-pass filtering the raw FCz signal at 95% of the Nyquist frequency (500 Hz) to isolate pulsatile transients. Peak detection was performed on each of these time series using the findpeaks MATLAB function with a critical interpeak interval of 400 ms. The union of peaks detected from these two methods was used to identify artifact locations. These timestamps were used to segment the sensors voltages into 200 ms time windows starting 50 ms before the stimulus pulse and ending 150 ms after the stimulus pulse (Fig. 2B). Time segments with motion and blink artifacts were removed using a single high amplitude cutoff of 200 µV and the remaining time segments were averaged to produce an EP at each sensor (Fig. 2C). Finally, sensor voltages were baseline corrected to the average voltage from −30 ms to 0, the start of the stimulus pulse.

Source Localization.

Source localization was performed on EEG data recorded during individual stimulation experiments using standardized low-resolution electromagnetic tomographic analysis (sLORETA) with a signal-to-noise ratio (SNR) of 3.0 and regularization parameter = 1/2*SNR. The fsaverage head model with corresponding 256-lead EEG lead locations was utilized to compute the inverse solution in each case. The Human Connectome Project (HCP) parcellation of the fsaverage brain was utilized to label voxels within specific anatomic regions (56). A representative source time course (STC) was computed for each anatomic region by averaging the amplitude at each voxel in the respective region. The source analysis was performed in python 3.11.5 using the following packages: mne 1.5.1, numpy 1.24.4, matplotlib 3.8.0, scipy 1.11.2, pandas 2.1.1, and seaborn 0.12.2 (57–62).

Statistical Analysis.

The left–right symmetry of EPs was quantified using the average time series correlation within 30 ms time windows (30 ± 15, 60 ± 15, 90 ± 15, and 120 ± 15 ms) for each left–right sensor pair in the 256-sensor EEG (Fig. 3C). The correlation between discrete voltages V of sensor pair i consisting of left sensor il and right sensor ir within window T is given as[2] ρi,T=corrVilT,VirT.

The grand average of correlations within this time window can thus be computed by averaging the individual correlation of each sensor pair i in the set of n pairs.[3] ρT=1n∑i=1nρi,T.

Homogeneity of variance and normality were assessed across data grouped by side of stimulation, time-window, and bundle recruitment group (63, 64). Both tests returned a P < 0.001, indicating unequal variance and non-normally distributed data. Due to these characteristics, we utilized Kruskal–Wallis tests at each time point, followed by Dwass–Steel–Critchlow–Fligner pairwise comparisons (65, 66) to assess difference in ρ between recruitment groups (i.e., CB-selective, nonselective, FM-selective). Centroids of positive and negative dipoles were computed by first spatially averaging the channels with the largest positive and negative EP amplitudes and their adjacent neighbors, interpolating signal for missing channels. These results were interpolated over a 200-by-200 grid of points mapped onto the topogram. Points with amplitudes in the top 10% and points with amplitudes in the bottom 10% were used to compute positive and negative dipole centroids, respectively.

Supplementary Material

Appendix 01 (PDF)

The DBS devices used in this study were donated by Medtronic, Inc. This work made use of the High-Performance Computing Resource in the Core Facility for Advanced Research Computing at Case Western Reserve University and the Duke Compute Cluster at Duke University. NIH grant R01MH102238 (C.C.M.). NIH grant UH3NS103550 (H.S.M., P.R.-P., and C.J.R.).

Author contributions

A.S., M.S.N., C.J.R., H.S.M., C.C.M., A.C.W., and B.H. designed research; A.S., M.S.N., A.V., M.O., J.D.-F., V.T., E.X., P.R.-P., A.C.W., and B.H. performed research; A.S., M.S.N., C.C.M., A.C.W., and B.H. analyzed data; and A.S., M.S.N., K.S.C., H.S.M., C.C.M., A.C.W., and B.H. wrote the paper.

Competing interests

C.C.M. is a paid consultant for Boston Scientific, Co., receives royalties from Hologram Consultants, Neuros Medical, Ceraxis, and Qr8 Health, and is a shareholder in the following companies: Hologram Consultants, Surgical Information Sciences, CereGate, Cardionomic, Enspire DBS. All other authors declare that they have no competing interests. H.S.M. receives consulting and IP licensing fees from Abbott Labs. K.S.C. is a paid consultant for Abbott Labs.

Data, Materials, and Software Availability

All code utilized for analysis is available on GitHub and can be found at this link: https://github.com/AndreasSeas/SCC_DBS_EP (67). Subjects’ medical images and EEG can be shared in a deidentified/anonymized form with a data use agreement upon request.

Supporting Information

This article is a PNAS Direct Submission.
==== Refs
1 H. S. Mayberg , Deep brain stimulation for treatment-resistant depression. Neuron 45 , 651–660 (2005).15748841
2 P. E. Holtzheimer , Subcallosal cingulate deep brain stimulation for treatment-resistant depression: A multisite, randomised, sham-controlled trial. Lancet Psychiatry 4 , 839–849 (2017).28988904
3 M. Sobstyl, A. Kupryjaniuk, M. Prokopienko, M. Rylski, Subcallosal cingulate cortex deep brain stimulation for treatment-resistant depression: A systematic review. Front. Neurol. 13 , 780481 (2022).35432155
4 W. M. Grill, A. N. Snyder, S. Miocinovic, Deep brain stimulation creates an informational lesion of the stimulated nucleus. Neuroreport 15 , 1137–1140 (2004).15129161
5 C. Hamani , Deep brain stimulation of the subcallosal cingulate gyrus for depression: Anatomical location of active contacts in clinical responders and a suggested guideline for targeting. J. Neurosurg. 111 , 1209–1215 (2009).19480538
6 J. L. Lujan , Tractography-activation models applied to subcallosal cingulate deep brain stimulation. Brain Stimul. 6 , 737–739 (2013).23602025
7 P. Riva-Posse , Defining critical white matter pathways mediating successful subcallosal cingulate deep brain stimulation for treatment-resistant depression. Biol. Psychiatry 76 , 963–969 (2014).24832866
8 P. Riva-Posse , A connectomic approach for subcallosal cingulate deep brain stimulation surgery: Prospective targeting in treatment-resistant depression. Mol. Psychiatry 23 , 843–849 (2018).28397839
9 B. Howell , Quantifying the axonal pathways directly stimulated in therapeutic subcallosal cingulate deep brain stimulation. Hum. Brain Mapp. 40 , 889–903 (2019).30311317
10 S. Alagapan , Cingulate dynamics track depression recovery with deep brain stimulation. Nature 622 , 130–138 (2023).37730990
11 C. R. Conner, T. M. Ellmore, T. A. Pieters, M. A. DiSano, N. Tandon, Variability of the relationship between electrophysiology and BOLD-fMRI across cortical regions in humans. J. Neurosci. 31 , 12855–12865 (2011).21900564
12 Y. Yamao , Clinical impact of intraoperative CCEP monitoring in evaluating the dorsal language white matter pathway. Hum. Brain Mapp. 38 , 1977–1991 (2017).28112455
13 C. J. Keller , Intrinsic functional architecture predicts electrically evoked responses in the human brain. Proc. Natl. Acad. Sci. U.S.A. 108 , 10308–10313 (2011).21636787
14 C. J. Keller , Mapping human brain networks with cortico-cortical evoked potentials. Philos. Trans. R. Soc. Lond. B, Biol. Sci. 369 , 20130528 (2014).25180306
15 P. A. Chouinard, T. Paus, What have we learned from “perturbing” the human cortical motor system with transcranial magnetic stimulation? Front. Hum. Neurosci. 4 , 173 (2010).21060721
16 R. Matsumoto , Functional connectivity in the human language system: A cortico-cortical evoked potential study. Brain 127 , 2316–2330 (2004).15269116
17 S. Ookawa , Frontal fibers connecting the superior frontal gyrus to broca area: A corticocortical evoked potential study. World Neurosurg. 107 , 239–248 (2017).28797973
18 J. Parvizi, S. Kastner, Human Intracranial EEG: Promises and limitations. Nat. Neurosci. 21 , 474–483 (2018).29507407
19 J. A. Adkinson , Imaging versus electrographic connectivity in human mood-related fronto-temporal networks. Brain Stimul. 15 , 554–565 (2022).35292403
20 B. Burle , Spatial and temporal resolutions of EEG: Is it really black and white? A scalp current density view. Int. J. Psychophysiol. 97 , 210–220 (2015).25979156
21 A. C. Waters , Test–retest reliability of a stimulation-locked evoked response to deep brain stimulation in subcallosal cingulate for treatment resistant depression. Hum. Brain Mapp. 39 , 4844–4856 (2018).30120851
22 S. An , High-resolution virtual brain modeling personalizes deep brain stimulation for treatment-resistant depression: Spatiotemporal response characteristics following stimulation of neural fiber pathways. NeuroImage 249 , 118848 (2021).34954330
23 K. Gunalan , Creating and parameterizing patient-specific deep brain stimulation pathway-activation models using the hyperdirect pathway as an example. PLoS ONE 12 , e0176132 (2017).28441410
24 B. Howell, C. C. McIntyre, Analyzing the tradeoff between electrical complexity and accuracy in patient-specific computational models of deep brain stimulation. J. Neural. Eng. 13 , 036023 (2016).27172137
25 B. Howell , “Connectomic predictive modeling guides selective perturbation of tracts in the subcallosal cingulate white matter” in 10th International IEEE/EMBS Conference on Neural Engineering (NER) (2021), pp. 271–274.
26 D. Momi , Perturbation of resting-state network nodes preferentially propagates to structurally rather than functionally connected regions. Sci. Rep. 11 , 12458 (2021).34127688
27 R. M. Fame, J. L. MacDonald, J. D. Macklis, Development, specification, and diversity of callosal projection neurons. Trends Neurosci. 34 , 41–50 (2011).21129791
28 M. Zarei , Functional anatomy of interhemispheric cortical connections in the human brain. J. Anat. 209 , 311–320 (2006).16928200
29 H.-J. Park , Corpus callosal connection mapping using cortical gray matter parcellation and DT-MRI. Hum. Brain Mapp. 29 , 503–516 (2006).
30 J.-E. Kim , Hong, diffusion tensor tractography analysis of the corpus callosum fibers in amyotrophic lateral sclerosis. J. Clin. Neurol. (Seoul, Korea) 10 , 249–256 (2014).
31 J. Cha , Whole brain network effects of subcallosal cingulate deep brain stimulation for treatment-resistant depression. Mol. Psychiatry. 28 , 1–9 (2023).36599928
32 M. Suzuki, M. E. Larkum, Dendritic calcium spikes are clearly detectable at the cortical surface. Nat. Commun. 8 , 276 (2017).28819259
33 S. Næss , Biophysically detailed forward modeling of the neural origin of EEG and MEG signals. NeuroImage 225 , 117467 (2021).33075556
34 H. Blumenfeld, D. A. McCormick, Corticothalamic inputs control the pattern of activity generated in thalamocortical networks. J. Neurosci. 20 , 5153–5162 (2000).10864972
35 D. K. Jones, K. F. Christiansen, R. J. Chapman, J. P. Aggleton, Distinct subdivisions of the cingulum bundle revealed by diffusion MRI fibre tracking: Implications for neuropsychological investigations. Neuropsychologia 51 , 67–78 (2013).23178227
36 Y. Wu, D. Sun, Y. Wang, Y. Wang, S. Ou, Segmentation of the cingulum bundle in the human brain: A new perspective based on DSI tractography and fiber dissection study. Front. Neuroanat. 10 , 84 (2016).27656132
37 S. R. Heilbronner, S. N. Haber, Frontal cortical and subcortical projections provide a basis for segmenting the cingulum bundle: Implications for neuroimaging and psychiatric disorders. J. Neurosci. 34 , 10041–10054 (2014).25057206
38 G. M. Innocenti, T. B. Dyrby, K. W. Andersen, E. M. Rouiller, R. Caminiti, The crossed projection to the striatum in two species of monkey and in humans: Behavioral and evolutionary significance. Cereb. Cortex 27 , 3217–3230 (2017).27282154
39 D. L. Clark , Tract-based analysis of target engagement by subcallosal cingulate deep brain stimulation for treatment resistant depression. Brain Stimul. 13 , 1094–1101 (2020).32417668
40 G. J. B. Elias , 3 T MRI of rapid brain activity changes driven by subcallosal cingulate deep brain stimulation. Brain 145 , 2214–2226 (2022).34919630
41 K. S. Choi, P. Riva-Posse, R. E. Gross, H. S. Mayberg, Mapping the “depression switch” during intraoperative testing of subcallosal cingulate deep brain stimulation. JAMA Neurol. 72 , 1252–1260 (2015).26408865
42 G. J. B. Elias , Local neuroanatomical and tract-based proxies of optimal subcallosal cingulate deep brain stimulation. Brain Stimul. 16 , 1259–1272 (2023).37611657
43 R. Hanajima , Interhemispheric facilitation of the hand motor area in humans. J. Physiol. 531 , 849–859 (2001).11251064
44 S. M. Smith , Advances in functional and structural MR image analysis and implementation as FSL. NeuroImage 23 , S208–S219 (2004).15501092
45 R. W. Cox, AFNI: Software for analysis and visualization of functional magnetic resonance neuroimages. Comput. Biomed. Res., Int. J. 29 , 162–173 (1996).
46 A. Husch, M. V. Petersen, P. Gemmar, J. Goncalves, F. Hertel, PaCER—A fully automated method for electrode trajectory and contact reconstruction in deep brain stimulation. Neuroimage Clin. 17 , 80–89 (2018).29062684
47 Z. Zhu , Individualized targeting is warranted in subcallosal cingulate gyrus deep brain stimulation for treatment-resistant depression: A tractography analysis. Hum. Brain Mapp. 44 , 4200–4421 (2023).37227015
48 P. A. Cooke , “Camino: Open-Source Diffusion-MRI Reconstruction and Processing” in 14th Scientific Meeting of the International Society for Magnetic Resonance in Medicine (Seattle, WA, USA, 2006), p. 2759.
49 S. Mori, B. J. Crain, V. P. Chacko, P. C. van Zijl, Three-dimensional tracking of axonal projections in the brain by magnetic resonance imaging. Ann. Neurol. 45 , 265–269 (1999).9989633
50 M. Fabri, C. Pierpaoli, P. Barbaresi, G. Polonara, Functional topography of the corpus callosum investigated by DTI and fMRI. World J. Radiol. 6 , 895–906 (2014).25550994
51 D. Liewald, R. Miller, N. Logothetis, H.-J. Wagner, A. Schüz, Distribution of axon diameters in cortical white matter: An electron-microscopic study on three human brains and a macaque. Biol. Cybern. 108 , 541–557 (2014).25142940
52 L. Firmin , Axon diameters and conduction velocities in the macaque pyramidal tract. J. Neurophysiol. 112 , 1229–1240 (2014).24872533
53 B. Howell, C. C. McIntyre, Role of soft-tissue heterogeneity in computational models of deep brain stimulation. Brain Stimul. 10 , 46–50 (2017).27720186
54 B. Howell, K. Gunalan, C. C. McIntyre, A driving-force predictor for estimating pathway activation in patient-specific models of deep brain stimulation. Neuromodulation 22 , 403–415 (2019).30775834
55 B. Howell, B. Huynh, W. M. Grill, Design and in vivo evaluation of more efficient and selective deep brain stimulation electrodes. J. Neural Eng. 12 , 046030 (2015).26170244
56 M. F. Glasser , A multi-modal parcellation of human cerebral cortex. Nature 536 , 171–178 (2016).27437579
57 A. Gramfort , MEG and EEG data analysis with MNE-Python. Front. Neurosci.-Switz. 7 , 267 (2013).
58 C. R. Harris , Array programming with NumPy. Nature 585 , 357–362 (2020).32939066
59 N. Ari, M. Ustazhanov, “Matplotlib in Python” in Proceedings of the 2014 11th International Conference on Electronics, Computer and Computation (ICECCO’14) (2014).
60 P. Virtanen ., SciPy 1.0.: Fundamental algorithms for scientific computing in Python. Nat. Methods 17 , 261–272 (2020).32015543
61 W. McKinney, “Data Structures for Statistical Computing in Python” in Proceedings of the 9th Python in Science Conference, 28 June-3 July 2010, S. van der Walt, M. Jarrod, Eds. (Austin, 2010), pp. 56–61.
62 M. Waskom, seaborn: Statistical data visualization. J. Open Source Software 6 , 3021 (2021).
63 M. B. Brown, A. B. Forsythe, Robust Tests for the Equality of Variances. J. Am. Stat. Assoc. 69 , 364–367 (1974).
64 S. S. Shapiro, M. B. Wilk, An analysis of variance test for normality (complete samples). Biometrika 52 , 591–611 (1965).
65 W. H. Kruskal, W. A. Wallis, Use of ranks in one-criterion variance analysis. J. Am. Stat. Assoc. 47 , 583–621 (1952).
66 D. E. Critchlow, M. A. Fligner, On distribution-free multiple comparisons in the one-way analysis of variance. Commun. Stat. Theory 20 , 127–139 (1991).
67 A. Seas, SCC_DBS_EP. Github. https://github.com/AndreasSeas/SCC_DBS_EP. Deposited 7 January 2024.
